Learning Transferable Adversarial Examples via Ghost Networks

Introduction

This repository contains the code for paper Learning Transferable Adversarial Examples via Ghost Networks. In this paper, we propose Ghost Networks to efficiently learn transferable adversarial examples. The key principle of ghost networks is to perturb an existing model, which potentially generates a huge set of diverse models. Those models are subsequently fused by longitudinal ensemble. Both steps almost require no extra time and space consumption. Experiment shows this method could consistently gain additional transferability for iteration-based method (such as I-FGSM and MI-FGSM).

Extension

To improve the transferability further, we

• attack multiple networks simultaneously (https://arxiv.org/abs/1611.02770).

Usage

Dependencies

• Anaconda
• Python3.6
• Tensorflow 1.10.0
• Tensorpack 0.9.0.1
• easydict
• scipy
• pillow

Here is a sample scrip to install Dependencies after you have Anaconda.

conda create -n python3 python=3.6
source activate python3
pip install --upgrade tensorflow-gpu
pip install --upgrade git+https://github.com/tensorpack/tensorpack.git
pip install easydict
conda install -c anaconda scipy
pip install pillow

Dataset and model checkpoints

We use images from ImageNet LSVRC 2012 Validation Set and resized them to 299x299. You can download the preprocessed images HERE if you accept the terms.

We use 6 clean trained models (Inception-{v3, v4}, Resnet-v2-{50, 101, 152}, Inception-Resnet-v2) and 3 ensemble adversarial trained models (ens3_inception_v3, ens4_inception_v3, ens_inception_resnet_v2). We original download them from here and here and then slightly modified the tensor name. You can download the modified checkpoints from HERE.

After download them, edit and use data/link_to_data.sh to build soft link data/checkpoints and data/val_data by

bash data/link_to_data.sh

We assign every network with an id, so that they can be shortly mentioned in one character. Here is a table to provide ids for each network. You can see line 58 to 62 of config.py for more details.

ID	0	1	2	3	4	5	6	7	8
Network Name	IncV3	IncV4	Res50	Res101	Res152	IncRes	Ens3IncV3	Ens3IncV4	EnsIncRes

Attack and Eval Examples

This section provide some examples, you can check config.py for more options.

Basic FGSM, I-FGSM, and MI-FGSM

FGSM attack inception_v3(ID=0) and evaluate success rate

bash pipeline.sh --exp FGSM --attack_network 0 --num_steps 1 --max_epsilon 8.0 --step_size 8.0 --GPU_ID 0

I-FGSM attack inception_v3(ID=0) and evaluate success rate

bash pipeline.sh --exp I-FGSM --attack_network 0 --GPU_ID 0

MI-FGSM attack inception_v3(ID=0) and evaluate success rate

bash pipeline.sh --exp MI-FGSM --attack_network 0 --momentum 1.0 --GPU_ID 0

Our proposed method

MI-FGSM attack inception_v3(ID=0) with dropout erosion (with the optimal keep_prob) and evaluate success rate

# these two line of scripts are same, since the optimal keep_prob for inception_v3 is 0.994
bash pipeline.sh --exp MI-FGSM-0.994 --attack_network 0 --momentum 1.0 --keep_prob 0.994 --GPU_ID 0
bash pipeline.sh --exp MI-FGSM-optimal --attack_network 0 --momentum 1.0 --optimal --GPU_ID 0

MI-FGSM attack resnet_v2_50(ID=2) with residual erosion (with the optimal Lambda) and evaluate success rate

# these two line of scripts are same, since the optimal Lambda for resnet_v2_50 is 0.22
bash pipeline.sh --exp MI-FGSM --attack_network 2 --momentum 1.0 --random_range 0.22 --GPU_ID 0
bash pipeline.sh --exp MI-FGSM --attack_network 2 --momentum 1.0 --optimal --GPU_ID 0

Attack multiple networks (Ensemble attack, Liu et al)

Simply put all network ids to the parameter of --attack_network, if your GPU memory is not enough, reduce the --batch_size.

For example, attack the ensemble of inception_v3 and inception_v4 with batch_size=2

bash pipeline.sh --exp FGSM --attack_network 01 --num_steps 1 --max_epsilon 8.0 --step_size 8.0 --batch_size 2 --GPU_ID 0

Trouble Shooting

Traceback (most recent call last):
  File "attack.py", line 5, in <module>
    from config import config as FLAGS
  File "/home/yingwei/lyw/mount_point/ghost-network/config.py", line 81, in <module>
    assert config.overwrite or config.skip, "{:s}".format(config.result_dir)
AssertionError: result/I-FGSM...

This Assertion Raise due to config.result_dir already exists, mainly due to you run the same scripts twice. You can either

1. remove that director
2. add option --overwrite to ignore this issue

If you find the code useful, please consider citing the following paper.

@article{li2018learning,
  title={Learning Transferable Adversarial Examples via Ghost Networks},
  author={Li, Yingwei and Bai, Song and Zhou, Yuyin and Xie, Cihang and Zhang, Zhishuai and Yuille, Alan},
  journal={arXiv preprint arXiv:1812.03413},
  year={2018}
}

If you encounter any problems or have any inquiries, please contact us at yingwei.li@jhu.edu.